Junda (David) Su

PhD student in Data Science at UC San Diego. PDF version

Education

PhD in Data Science, La Jolla, CA

Sep 2025 - Present

Bachelor of Science in Computer Science, Houston, TX

Aug 2021 - May 2025

Mentor: Prof. Anshumali Shrivastava

Aug 2024 - Oct 2024

Proposed SpaLLM, a fine-tuning paradigm on compressed LLMs using parameter-sharing algorithms.
Benchmarked across LLaMA-2-7B/13B, LLaMA-3-8B and 70B, achieving up to 3x inference speedup over adapter-based compressive fine-tuning.

Mentor: Prof. Zhaozhuo Xu

Dec 2023 - Jun 2024

Proposed SpartanServe, a system for fast concurrent LLM adapter serving with structurally sparse adapters.
Used Triton kernels and CUDA graphs to accelerate batched matrix multiplication.
Achieved 2.12x speedup over S-LoRA when serving 96 adapters on a single A100 (40GB).

Mentor: Prof. Vladimir Braverman

Aug 2023 - Oct 2023

Developed a CNN + BiLSTM model for arrhythmia classification on real-world wearable ECG data.
Achieved 95% average accuracy on binary and 88% on multi-label classification.

Mentor: Prof. Robert Waterland

Aug 2022 - Dec 2022

Built a sequence-sampling API for whole-genome DNA methylation analysis software.
Parallelized the pipeline on a Linux cluster for a 20x data processing speedup.

Tianyi Zhang*, Junda Su*, Aditya Desai, Oscar Wu, Zhaozhuo Xu, Anshumali Shrivastava. Sketch to Adapt: Fine-Tunable Sketches for Efficient LLM Adaptation. ICML 2025.
Junda Su, Zirui Liu, Zeju Qiu, Weiyang Liu, Zhaozhuo Xu. In Defense of Structural Sparse Adapters for Concurrent LLM Serving. Findings of EMNLP 2024.
Guangyao Zheng, Sunghan Lee, Jeonghwan Koh, Khushbu Pahwa, Haoran Li, Zicheng Xu, Haiming Sun, Junda Su, Sung Pil Cho, Sung Il Im, In cheol Jeong, Vladimir Braverman. Hierarchical Deep Learning for Autonomous Multilabel Arrhythmia Detection and Classification on Real-world Wearable ECG Data. Digital Health 2024.

* indicates equal contribution

Technology Advancement Program Intern, Houston, TX

May 2023 - Aug 2023

Designed quote submission and retrieval REST APIs for an insurance website.
Used AWS API Gateway, MongoDB, and PostgreSQL for traffic scaling and data management.
Led daily standups and biweekly sprint planning.

Teaching Assistant, Rice University